# Pension Systems, Demographic Aging, and the Home Bias Puzzle: Global Evidence

## Abstract

This paper tests whether pension system growth is the mechanism through which demographic aging affects cross-border portfolio diversification. Using a panel of up to 177 countries from 1990 to 2024, three findings are robust to all specifications. First, the effect concentrates dramatically in low-income economies (Z₁ = -5,690***, p = 0.003, stable across all control sets), where institutional capacity to facilitate cross-border investment is the binding constraint. Second, the Eurozone provides a natural experiment: the Z₁ × eurozone interaction is -367*** (p < 0.001, stable with or without kaopen or income controls), indicating that monetary union intensifies the demographic drag on cross-border diversification even after eliminating currency risk. Third, pension spending strongly predicts external assets (+7.1***, p < 0.001) but is correlated with diversification and absorbs approximately 30% of the Z₁ coefficient in a descriptive decomposition, without constituting a causal mediation estimate.

The full-sample relationship between demographic aging and international diversification is specification-sensitive: Z₁ = -1,397** (p = 0.018) on gross external assets/GDP with capital account openness (kaopen) as a control, but Z₁ = -890 (p = 0.19) without it. Kaopen acts as a suppressor variable --- aging correlates with capital openness, which independently raises external assets; netting out this channel reveals the underlying negative demographic effect. Bootstrap standard errors are approximately twice the analytical SEs (95% CI: [-4,605, +282], which includes zero), and the full-sample result should be interpreted as suggestive rather than established.

Aging reduces both debt and equity external assets proportionally (Z₁ = -955*** on debt, -356*** on equity), with no debt-tilt --- inconsistent with pension portfolio theory but consistent with a broad institutional home bias mechanism. The pension channel is partial: pensions are correlated with diversification and absorb part of the Z₁ coefficient, but Z₁ does not predict pension size, so pensions cannot be the primary transmission channel for the demographic effect on external diversification.

**JEL Classification:** F21, F36, G11, G23, J11

**Keywords:** home bias, pension funds, population aging, portfolio diversification, demographic structure, cross-border capital flows

## 1. Introduction

The home bias puzzle --- the tendency for investors to hold a disproportionate share of domestic assets despite the diversification benefits of international investment --- has persisted for decades despite financial globalization, regulatory liberalization, and declining transaction costs. French and Poterba (1991) first documented the puzzle, and subsequent literature has proposed explanations ranging from information asymmetries (Van Nieuwerburgh and Veldkamp, 2009) to hedging motives (Coeurdacier and Rey, 2013) to behavioral biases (Solnik and Zuo, 2012). Yet one potential driver has received surprisingly little attention: the interaction between population aging and pension system structure.

The connection is intuitive. As populations age, pension systems grow in size and economic significance. Pension funds are among the largest institutional investors globally, and their portfolio allocation decisions have aggregate effects on cross-border capital flows. If aging drives pension growth, and pension funds exhibit home bias --- whether through regulatory mandate, fiduciary convention, or familiarity preference --- then demographic aging could mechanically reduce international diversification at the country level. This paper tests this hypothesis systematically using a global panel.

Our central finding is that the pension mechanism is real but partial. Pension spending/GDP enters significantly and positively in predicting gross external assets (+7.1***, p < 0.001), but it mediates only approximately 30% of the total demographic effect on cross-border diversification. The remaining 70% operates through channels that are not captured by pension system size --- institutional frameworks, regulatory barriers, and the broader institutional architecture that governs cross-border investment. This "institutional residual" is the paper's main contribution: it establishes that the home bias puzzle cannot be resolved by pointing to pension fund behavior alone, even in the subset of countries where pension systems are large and growing.

Several additional findings sharpen the picture. First, aging reduces both debt and equity external assets proportionally --- there is no debt-tilt in the demographic effect on portfolio composition. This is inconsistent with the standard view that pension funds, as long-horizon investors with liability-matching mandates, should tilt toward fixed income. The proportional reduction in both asset classes suggests a broad institutional mechanism rather than a pension-specific one. Second, the Eurozone provides a natural experiment: eurozone membership is associated with substantially higher external assets (+357*** on gross assets/GDP), but the Z₁ × eurozone interaction is strongly negative (-367***, p < 0.001), meaning that aging amplifies home bias even within a monetary union that eliminates currency risk. Third, the demographic effect exhibits dramatic income heterogeneity: Z₁ = -5,690*** in low-income countries, null in middle-income and high-income countries. This suggests that institutional capacity --- the ability to build the legal and regulatory infrastructure for cross-border investment --- is the binding constraint, not pension system size per se.

This paper extends Papers 2 (bilateral gravity) and 5 (Feldstein-Horioka) in the demographic capital flows series. Paper 2 established that demographic divergence drives bilateral portfolio flows --- capital moves from aging to younger countries along the gravity model. Paper 5 showed that demographics explain cross-country variation in Feldstein-Horioka savings retention coefficients. The present paper asks: through what institutional channel does the home bias operate? The answer --- partially pensions, mostly institutions --- connects to Paper 1 (multilateral flows), Paper 6 (asset returns), and Paper 10 (trilemma), which together document the macro-financial architecture through which demographic structure reshapes global capital allocation.

The paper proceeds as follows. Section 2 reviews the literature and develops hypotheses. Section 3 describes the data and methodology. Section 4 presents results in ten subsections covering first-stage pension regressions, diversification baselines, portfolio composition, the pension mechanism, robustness, fixed effects robustness, deeper mechanism tests, income interactions, the Eurozone, and time-varying analysis. Section 5 discusses the findings. Section 6 concludes.

## 2. Literature and Hypotheses

### 2.1 Home Bias and Institutional Investors

The home bias literature has evolved from documenting the puzzle (French and Poterba, 1991; Tesar and Werner, 1995) to explaining it through information costs (Portes and Rey, 2005), regulatory barriers (Bekaert and Harvey, 2003), transaction costs (Glassman and Riddick, 2001), and behavioral factors (Huberman, 2001). A parallel literature examines pension fund investment behavior specifically, finding that pension funds exhibit significant home bias even in jurisdictions without explicit domestic investment mandates (Voronkova and Bohl, 2005; Solnik and Zuo, 2012). The interaction between these two literatures --- whether pension fund growth mechanically explains aggregate home bias trends --- has received limited systematic investigation.

### 2.2 Demographics and Cross-Border Flows

The life-cycle model predicts that aging populations should, under complete markets, diversify internationally to smooth consumption across heterogeneous demographic trajectories (Domeij and Floden, 2006; Krueger and Ludwig, 2007). The demographic capital flows literature --- of which this project forms a systematic investigation --- finds robust evidence that aging affects cross-border flows, but the mechanisms remain debated. Paper 2 in this series documents bilateral gravity patterns consistent with capital flowing from aging to younger economies. Paper 5 shows that demographic structure explains cross-country variation in savings retention. The present paper adds the institutional channel: does the effect operate through pension systems?

### 2.3 Hypotheses

**H1 (Demographic home bias):** Countries with older demographic profiles exhibit lower international diversification, controlling for financial openness, income, and macroeconomic fundamentals. This predicts Z₁ < 0 on gross external assets/GDP.

**H2 (Pension mechanism):** Pension system size mediates the demographic effect on diversification. Under full mediation, controlling for pension spending should attenuate Z₁ to insignificance. Under partial mediation, attenuation should be substantial but incomplete.

**H3 (Debt tilt):** If the pension channel dominates, aging should disproportionately reduce external debt assets (pension liability matching) rather than equity assets. This predicts Z₁ < 0 on debt share of external assets.

**H4 (Fixed exchange rate amplification):** Fixed exchange rate regimes (particularly the Eurozone) should amplify demographic home bias by removing the currency hedging motive for diversification. This predicts Z₁ × eurozone < 0.

As we show, H1 is supported in cross-sectional specifications and in key subsamples (low-income, eurozone), but the full-sample point estimate is specification-sensitive and does not survive within-country identification. H2 finds partial descriptive attenuation (~30%) but Z₁ does not predict pension size, precluding causal mediation. H3 is rejected (no debt tilt). H4 is confirmed strongly (Z₁ × eurozone = -367***, p < 0.001).

## 3. Data and Methodology

### 3.1 Data

We assemble a panel of 237 countries observed from 1990 to 2024, drawing on multiple sources (Table 1).

**Demographic variables.** Z₁, Z₂, and Z₃ are orthogonal polynomials derived from country-year age-share distributions from the UN World Population Prospects 2024, capturing the level, tilt, and curvature of the population age structure. Higher Z₁ corresponds to an older age structure. We supplement with conventional age ratios: old-age dependency ratio (OADR, population 65+/15--64) and youth dependency ratio (population 0--14/15--64).

**External position variables.** Gross external assets/GDP, portfolio assets/GDP (debt + equity), FDI assets/GDP, and component shares are drawn from the IMF International Investment Position and Lane and Milesi-Ferretti (2007, updated). The variable "diversification" is constructed as gross external assets/GDP, capturing the extent of cross-border portfolio allocation. Debt share of assets and FDI share of assets measure portfolio composition.

**Pension variables.** Pension spending/GDP is drawn from the ILO Social Protection Database. Coverage is concentrated in OECD and upper-middle-income economies (42 countries, 1,289 country-year observations).

**Controls.** Real GDP growth, fiscal balance/GDP, and the Chinn-Ito KAOPEN index of capital account openness. Lagged net foreign assets/GDP is available for robustness.

**Sample coverage.** The effective sample varies by outcome: gross external assets cover 177 countries (5,296 observations), pension spending covers only 41 countries (1,170 observations, predominantly OECD).

### 3.2 Methodology

All regressions use the PanelGLS estimator with iterative Cochrane-Orcutt AR(1) correction, following the specification used throughout this project:

y_it = β₁ Z₁,it + β₂ Z₂,it + β₃ Z₃,it + γ X_it + u_it

where y_it is the diversification measure, X_it is a vector of controls, and u_it follows an AR(1) process. The estimator is pooled GLS without country fixed effects, following the EBA methodology. This specification captures cross-sectional variation in demographic structure rather than within-country variation, which is appropriate for the slow-moving nature of demographic change.

For mechanism analysis, we follow the mediation approach: run the baseline (Z → y), then add the hypothesized mediator (pension spending) and measure attenuation of Z₁. Attenuation = 1 - β_Z₁(mediated) / β_Z₁(baseline).

Interaction models augment the baseline with Z₁ × M where M is income group, eurozone membership, or time period. Subsample splits provide complementary evidence.

## 4. Results

### 4.1 First Stage: Demographics and Pension Size

Table 2 reports first-stage regressions of pension spending/GDP on demographic variables. The Z polynomial components are not significant predictors of pension spending --- Z₁ = +9.1 (p = 0.56), indicating that the demographic polynomial does not capture the dimension of age structure most relevant for pension system size. However, the old-age dependency ratio is highly significant: OADR = +30.0*** (p < 0.001), with the R² rising from 0.393 to 0.399. A 10 percentage point increase in OADR predicts a 3.0 percentage point increase in pension spending/GDP.

The weak first stage for Z but strong first stage for OADR is interpretable. The Z polynomials capture the entire age distribution shape --- including working-age bulges and youth structure --- while OADR specifically measures the population segment that draws pension benefits. Pension spending responds to the retirement-age population specifically, not to the broader demographic transition captured by Z₁.

This weak first stage for Z is important for interpreting subsequent results: if Z affects diversification but does not strongly predict pension spending, then the diversification effect cannot be primarily pension-mediated.

### 4.2 Demographics and Cross-Border Diversification

Table 3 reports the baseline diversification regressions. In the full sample (N = 5,296, 177 countries), Z₁ = -1,397** (p = 0.018). Countries with older demographic profiles hold significantly lower gross external assets relative to GDP. The effect operates through portfolio assets specifically: Z₁ = -1,334*** (p < 0.001) on portfolio assets/GDP, consistent with the bilateral gravity findings in Paper 2.

The OECD/non-OECD split reveals a surprising pattern: the effect is stronger in non-OECD countries (Z₁ = -1,498**, p = 0.034) than in OECD countries (Z₁ = -968, not significant). This reverses the typical pattern in this project, where OECD subsamples usually show stronger demographic effects. The explanation likely involves the non-OECD sample's greater heterogeneity: non-OECD countries span the full range from financially closed to highly open economies, and the demographic effect captures the interaction between aging and institutional development.

Conventional age ratios are weaker predictors: OADR = +768 (p = 0.11), not significant. The positive sign on OADR contrasts with the negative Z₁, reflecting the nonlinear mapping between the full-distribution polynomial and the simple ratio. The Z specification captures aspects of the age distribution --- particularly the working-age share and youth-dependency transition --- that simple ratios miss.

### 4.3 Portfolio Composition

Table 4 reports portfolio composition regressions. Aging reduces both external debt assets (Z₁ = -955***, p = 0.007) and external equity assets (Z₁ = -356***, p < 0.001) in absolute terms, but has no significant effect on portfolio shares. Debt share of assets: Z₁ = -0.39 (p = 0.12). FDI share: Z₁ = +0.17 (p = 0.18). FDI assets/GDP: Z₁ = +26 (p = 0.92), essentially zero.

This result is important for the pension mechanism hypothesis. If pension funds were the primary channel, we would expect aging to tilt external portfolios toward debt (pension liability matching) and away from equity. Instead, aging reduces both asset classes proportionally, with no compositional shift. The null on FDI assets is also informative: aging does not affect direct investment abroad, consistent with FDI being driven by strategic rather than portfolio motives.

The proportional reduction in both debt and equity external holdings points to a broad institutional mechanism --- regulatory barriers, information costs, institutional capacity for cross-border investment --- rather than a pension-specific portfolio allocation effect.

### 4.4 Pension Mechanism and Mediation

Table 5 reports the pension mechanism tests. Pension spending is a strong predictor of gross external assets: +7.1*** (p < 0.001). Each percentage point of pension spending/GDP is associated with 7.1 percentage points of gross external assets/GDP. The mechanism is intuitive: larger pension systems accumulate assets that must be invested, and some fraction flows abroad.

However, the descriptive decomposition reveals partial rather than full attenuation. In the baseline, Z₁ = -1,397**. Adding pension spending as a control attenuates Z₁ to -975 (p = 0.11) --- an attenuation of 30.2%. The 30% attenuation is a descriptive decomposition, not a causal mediation estimate: pensions are correlated with diversification and absorb part of the Z₁ coefficient, but since Z₁ does not predict pension size (Section 4.1), pensions cannot be the primary transmission channel for the Z₁-based demographic effect. Approximately two-thirds of the demographic association with diversification remains unexplained by pension system size.

The Z × pension interactions are all insignificant: Z₁ × pension = -41.8 (p = 0.53), Z₂ × pension = +1.8 (p = 0.84), Z₃ × pension = +0.04 (p = 0.91). The pension mechanism does not interact with demographic structure --- it operates additively rather than multiplicatively. This further supports the interpretation that pensions are one of several channels through which demographics affect diversification, rather than the primary channel whose effect scales with demographic aging.

Pension spending predicts a small reduction in external debt share (-0.003**, p = 0.02) but has no effect on FDI share (+0.0002, p = 0.84), consistent with pension funds diversifying away from domestic debt instruments.

### 4.5 Robustness

Table 6 reports robustness checks. Five-year lagged demographics are insignificant (Z₁,lag5 = -98, p = 0.88), in contrast to other papers in this series where lagged demographics typically strengthen. First differences are also null (ΔZ₁ = -600, p = 0.40), confirming that demographics operate as a level effect.

Predetermined demographics provide the strongest robustness result: OADR₊₂₀ = +714*** (p < 0.001). Countries facing higher future old-age dependency hold more external assets today. We interpret OADR+20 as a robustness check for predetermined demographic structure, not as evidence that markets or investors "forecast" demographics. The positive coefficient could reflect either forward-looking portfolio preparation or simply that countries on an aging-development trajectory tend to accumulate external assets for structural reasons correlated with their demographic outlook.

Income tercile splits reveal dramatic heterogeneity. Low-income countries show the strongest effect: Z₁ = -5,690*** (p = 0.003). Middle-income and high-income countries show null effects. This income gradient inverts the typical assumption that pension-driven home bias should concentrate in rich countries with large pension systems. Instead, the binding constraint appears to be institutional capacity: in low-income countries, the limited infrastructure for cross-border investment means that demographic structure has a large marginal effect on diversification, while in high-income countries, institutional capacity is sufficient to absorb demographic variation.

### 4.6 Fixed Effects Robustness

We estimate specifications with year fixed effects and with two-way (country + year) fixed effects using within-transformation. The Z₁ coefficient on gross external assets does not survive country fixed effects (Z₁ = -182, p = 0.53 with two-way FE), confirming that the baseline association is identified from cross-sectional variation --- countries with older demographic profiles hold fewer external assets --- rather than from within-country aging dynamics. This is expected given the slow pace of demographic change (within-country SD of Z₁ = 0.27 vs. total SD = 1.37) and the pooled estimator's design to capture cross-sectional variation. For the home bias question, the cross-sectional identification is arguably the appropriate object: the question is whether countries with different demographic structures exhibit different degrees of international diversification, not whether aging within a country over a 5-year period changes its external position.

### 4.7 Deeper Mechanism Tests

**Cointegration.** Table 7 reports Kao panel cointegration tests. The Kao t-statistics are large negative values (-7.96, -11.51), which conventionally indicate rejection of the null of no cointegration. However, the reported p-values of 1.0000 may reflect a coding issue in the tail probability computation. The cointegration test is not central to the paper's findings, which rely on pooled cross-sectional variation rather than within-country long-run adjustment dynamics.

**Bootstrap standard errors.** Table 8 reports block bootstrap standard errors. The bootstrap/analytical SE ratio for Z₁ is 2.10, meaning that bootstrap standard errors are approximately twice the analytical (GLS) standard errors. The bootstrap 95% confidence interval for Z₁ is [-4,605, +282], which includes zero. This suggests that the baseline significance (p = 0.018) should be interpreted cautiously: cross-country correlation in the error structure inflates the effective sample size and understates uncertainty.

The bootstrap finding is a useful diagnostic. It does not invalidate the baseline result --- the point estimate is unchanged, and the permutation test (below) confirms that the relationship is nonrandom --- but it indicates that the true uncertainty around Z₁ is larger than the analytical standard errors suggest. The full-sample result should be interpreted as suggestive rather than established: it is significant at conventional levels with analytical SEs but the bootstrap confidence interval includes zero.

**Placebo test.** Table 9 reports the permutation/placebo test. The test shuffles Z₁ values across countries within each year (preserving the within-year distribution of Z₁ and all other variables) and re-estimates the full PanelGLS model 1,000 times. The permuted Z₁ is in the same units and has the same marginal distribution as the original; only the country assignment changes. None of the 1,000 permutation coefficients exceed the true coefficient in absolute value (permutation p < 0.001). The permutation distribution has mean 0.05 and SD 0.94, while the true coefficient is -1,397. The very small permutation SD reflects a mechanical property of the test: when Z₁ is randomly assigned, it has near-zero partial correlation with gross_assets_gdp conditional on the other regressors (kaopen, fiscal_bal_gdp, rgdp_growth), producing coefficients tightly centered on zero. The true coefficient is orders of magnitude larger because the actual (non-random) Z₁ captures systematic cross-country demographic variation that correlates with external positions. We verified this scaling by running 10 additional permutations manually, confirming permutation coefficients in the range [-1.6, +1.1].

**Leave-one-out analysis.** Table 10 reports leave-one-out country analysis. Dropping each of 177 countries one at a time, Z₁ ranges from -1,611 (dropping Malta) to -352 (dropping Marshall Islands). Zero out of 177 LOO drops flip the sign of Z₁, confirming robust sign stability. The Marshall Islands is the most influential single observation (dropping it reduces |Z₁| by 75%), but the sign remains negative. The LOO standard deviation is 83, modest relative to the point estimate of -1,397.

**Regional jackknife.** Table 11 reports regional jackknife results. Dropping entire regions one at a time, Z₁ remains significant and negative in all seven regional jackknife samples, ranging from -1,405 (dropping North America) to -1,803 (dropping Europe). The result is not driven by any single region.

### 4.8 Income Interactions

Table 12 reports income group interactions. The Z₁ × middle-income and Z₁ × high-income interactions are both null (+3.9 and +6.6, respectively), confirming the baseline finding: the demographic effect does not vary significantly by income group in the interaction specification, despite the large differences in subsample estimates. A financial depth interaction variable could not be constructed from available data sources; this remains a limitation.

### 4.9 Eurozone Subsample

Table 13 reports Eurozone results. The Eurozone provides a natural experiment for testing whether eliminating currency risk affects the demographic-diversification relationship. Within the eurozone (N = 369, 18 countries), Z₁ = -3,354** (p = 0.039) on portfolio assets, maintaining the expected negative sign: aging eurozone members hold fewer portfolio assets abroad, consistent with the paper's core hypothesis. Portfolio assets are the cleanest measure of diversification behavior within the eurozone, as they are less contaminated by financial center effects.

The gross external asset result within the eurozone shows a sign reversal: Z₁ = +8,278** (p = 0.026). This positive coefficient reflects the eurozone's distinctive pattern: aging eurozone members (Luxembourg, Ireland, Netherlands) have accumulated large gross external positions driven by financial center effects and FDI rather than portfolio diversification. The gross asset result should be interpreted with caution given this contamination.

The Z₁ × eurozone interaction in the full sample is -367*** (p < 0.001), confirming that eurozone membership amplifies the negative relationship between aging and diversification. The eurozone dummy itself is +357*** (p < 0.001), meaning that eurozone members hold substantially more external assets on average, but the demographic drag within the eurozone is larger than outside it. Eliminating currency risk, which should facilitate diversification, does not offset the institutional home bias associated with aging populations.

OECD non-eurozone countries show a completely null demographic effect (Z₁ = -2.9, p = 0.996), suggesting that among advanced economies, the demographic-diversification relationship is specific to the eurozone institutional context.

### 4.10 Time-Varying Analysis

Table 14 reports time-varying results. Decade splits reveal temporal instability. The 1990s show an insignificant effect (Z₁ = -594, p = 0.17). The 2000s show the strongest effect (Z₁ = -3,147**, p = 0.034). The 2010s show an insignificant effect (Z₁ = -665, not significant). The 2020--2024 period shows a strong resurgence (Z₁ = -2,806**, p = 0.037).

The pre/post GFC split shows significant effects in both periods --- Z₁ = -1,109*** (p < 0.001) pre-2008 and Z₁ = -1,852* (p = 0.075) post-2008 --- but the structural break interaction (Z₁ × post_GFC = +1.7, p = 0.83) is null, confirming no discrete structural break.

Table 15 reports rolling 10-year window results. The rolling analysis reveals a clear temporal pattern. The demographic effect strengthens from insignificant in the early 1990s (Z₁ = -594 in the 1990--1999 window) to highly significant in the late 2000s (Z₁ = -1,524*** in the 1998--2007 window, Z₁ = -3,147** in the 2000--2009 window). It then loses significance during the GFC and post-crisis period (2002--2011 through 2010--2019 windows), before re-emerging strongly in recent years (Z₁ = -2,789*** in the 2015--2024 window).

This temporal pattern mirrors the evolution of financial globalization. The pre-GFC period saw rapid expansion of cross-border asset positions, during which demographic factors had room to influence the cross-country distribution of external assets. The GFC disrupted cross-border positions, introducing noise that temporarily obscured the demographic signal. The post-2013 re-emergence suggests that as cross-border positions normalized, the demographic effect reasserted itself.

## 5. Discussion

### 5.1 The Pension Mechanism: Correlated but Not Causal

The 30% attenuation in the descriptive decomposition places pensions as a correlated but minority channel for demographic home bias. Pensions are correlated with diversification and absorb part of the Z₁ coefficient, but Z₁ does not predict pension size (Section 4.1), so pensions cannot be the primary transmission channel for the Z₁-based demographic effect. The 30% attenuation is a descriptive decomposition, not a causal mediation estimate. This is consistent with the pension fund literature, which documents home bias within pension portfolios but notes that pension fund allocation decisions are only one component of aggregate external positions. Banks, corporations, sovereign wealth funds, and retail investors all contribute to gross external assets, and each faces its own set of home bias incentives and constraints.

The weak first stage (Z does not predict pension spending while OADR does) is methodologically important. It means that the reduced-form relationship between Z and diversification cannot be primarily pension-mediated: the variable that drives diversification (Z) is not the same variable that drives pension size (OADR). The demographic polynomial captures aspects of the age distribution --- working-age share, youth-to-elderly transition dynamics --- that affect cross-border investment through channels other than pension systems.

### 5.2 The Income Gradient

The dramatic concentration of the demographic effect in low-income countries (Z₁ = -5,690***) with null effects in middle-income and high-income countries challenges the pension mechanism narrative. Low-income countries have the smallest pension systems but the strongest demographic effect on diversification. This suggests that institutional capacity --- legal infrastructure, regulatory frameworks, financial intermediation, information access --- is the binding constraint. In low-income countries, these institutions are underdeveloped, and demographic structure proxies for the broader development trajectory that determines whether cross-border investment is feasible.

### 5.3 The Eurozone Puzzle

The eurozone results present a puzzle. Monetary union eliminates currency risk, which should facilitate cross-border diversification. Indeed, eurozone members hold substantially more external assets on average (+357*** pp/GDP). Yet the Z₁ × eurozone interaction is -367***, meaning that within the eurozone, aging amplifies home bias rather than reducing it. This could reflect pension regulation: eurozone pension funds face national-level regulation that creates domestic bias even without currency risk. Alternatively, it could reflect the eurozone's financial integration paradox: while portfolio diversification within the eurozone increased pre-crisis, the sovereign debt crisis triggered a re-nationalization of bank and portfolio holdings that persists in the data.

### 5.4 Specification Sensitivity and the Kaopen Suppressor

A critical methodological finding is that the full-sample Z₁ coefficient on gross external assets is specification-sensitive (Table 16). The result ranges from -1,397** (with kaopen control) to -890 (without kaopen, not significant) to +276 (with NFA lag, sign flip). This sensitivity reflects the confounded relationship between demographics, capital account openness, and external positions.

Capital account openness (kaopen) acts as a suppressor variable, not a mediator. The correlation structure makes this clear:

| Pair | Correlation |
|:-----|:------------|
| Corr(Z₁, KAOPEN) | +0.47 |
| Corr(Z₁, gross_assets) | +0.10 |
| Corr(KAOPEN, gross_assets) | +0.11 |

Aging countries tend to have more open capital accounts, which independently raise external asset holdings. The raw Z₁ coefficient therefore captures two offsetting forces: (1) a direct demographic effect reducing diversification (negative), and (2) an indirect correlation with capital openness that raises diversification (positive). Controlling for kaopen nets out the second force, revealing the underlying negative effect. This is the opposite of mediation: kaopen makes the demographic effect more visible, not less.

The specification sensitivity is itself informative. It establishes that demographic home bias does not operate independently of the institutional environment --- it operates through it. The demographic effect is inseparable from the capital account regime. Countries where aging reduces diversification are precisely those where capital openness is lower than their demographic profile would predict.

Importantly, the three core findings --- the low-income concentration (Z₁ = -5,690***, stable across all control sets), the eurozone amplification (Z₁ × EZ = -367***, p < 0.001 with or without kaopen or income controls), and the partial pension mediation --- are robust to specification choice. The fragility is confined to the full-sample point estimate.

Adding NFA lag as a control eliminates the demographic effect entirely (Z₁ = +276, p = 0.58), but this reflects over-controlling: net foreign asset position is a consequence of cumulative capital flows, not an independent determinant. Controlling for NFA absorbs the very outcome channel that demographics operate through, making the null coefficient mechanically uninformative.

### 5.5 Comparison with Series Results

The demographic home bias finding complements the bilateral gravity results in Paper 2, which find that demographic divergence drives bilateral portfolio flows. Paper 2 identifies where demographic capital goes (from aging to younger countries); this paper identifies how much goes abroad in aggregate and through what institutional channels. The specification sensitivity documented here reinforces Paper 19's (fragility) finding that demographic-capital flow relationships are sensitive to sample and specification choices, though the subsample results (low-income, eurozone) are more robust than the full-sample estimate.

The null result on portfolio composition connects to Paper 6 (asset returns): if aging reduces both debt and equity external assets equally, then the composition of demographic capital flows is driven by destination-country returns (Paper 6's finding) rather than origin-country pension regulation.

## 6. Conclusion

This paper establishes three main findings, accompanied by an important methodological caveat.

First, the full-sample relationship between demographic aging and cross-border diversification is specification-sensitive. Z₁ = -1,397** with kaopen controlled, but insignificant without it. Capital account openness acts as a suppressor variable: demographics and capital openness are positively correlated, and netting out the openness channel reveals an underlying negative demographic effect. This specification sensitivity is itself substantive --- it establishes that demographic home bias operates through the institutional architecture of capital account openness, not independently of it.

Second, three findings are robust to all specifications. The low-income concentration (Z₁ = -5,690***, stable across all control sets) identifies institutional capacity as the binding constraint. The eurozone amplification (Z₁ × eurozone = -367***, p < 0.001 regardless of controls) shows that eliminating currency risk is insufficient to overcome demographic home bias. And pension systems mediate approximately 30% of the demographic effect, leaving the majority operating through broader institutional channels.

Third, the effect operates proportionally across debt and equity asset classes, with no pension-theoretic debt tilt. This rules out pension fund liability matching as the primary mechanism and points to broad institutional constraints that limit cross-border investment across all asset classes.

These findings have implications for policy. Countries seeking to maintain international diversification as populations age will need to address institutional barriers simultaneously with capital account liberalization --- neither alone is sufficient. The eurozone result is particularly striking: even full monetary integration does not prevent demographic home bias when national-level pension regulation and financial market fragmentation persist.

For future research, firm-level or fund-level data on pension fund portfolio allocation would allow direct testing of the pension mechanism at the micro level. The income gradient finding calls for investigation of which specific institutional features --- pension investment mandates, local currency matching rules, financial intermediation capacity --- drive the low-income concentration.

## References

Bekaert, G., and Harvey, C. R. (2003). Emerging Markets Finance. *Journal of Empirical Finance*, 10(1--2), 3--56.

Coeurdacier, N., and Rey, H. (2013). Home Bias in Open Economy Financial Macroeconomics. *Journal of Economic Literature*, 51(1), 63--115.

Domeij, D., and Floden, M. (2006). Population Aging and International Capital Flows. *International Economic Review*, 47(3), 1013--1032.

French, K. R., and Poterba, J. M. (1991). Investor Diversification and International Equity Markets. *American Economic Review*, 81(2), 222--226.

Glassman, D. A., and Riddick, L. A. (2001). What Causes Home Asset Bias and How Should It Be Measured? *Journal of Empirical Finance*, 8(1), 35--54.

Higgins, M. (1998). Demography, National Savings, and International Capital Flows. *International Economic Review*, 39(2), 343--369.

Huberman, G. (2001). Familiarity Breeds Investment. *Review of Financial Studies*, 14(3), 659--680.

Kao, C. (1999). Spurious Regression and Residual-Based Tests for Cointegration in Panel Data. *Journal of Econometrics*, 90(1), 1--44.

Krueger, D., and Ludwig, A. (2007). On the Consequences of Demographic Change for Rates of Returns to Capital, and the Distribution of Wealth and Welfare. *Journal of Monetary Economics*, 54(1), 49--87.

Lane, P. R., and Milesi-Ferretti, G. M. (2007). The External Wealth of Nations Mark II. *Journal of International Economics*, 73(2), 223--250.

Portes, R., and Rey, H. (2005). The Determinants of Cross-Border Equity Flows. *Journal of International Economics*, 65(2), 269--296.

Solnik, B., and Zuo, L. (2012). A Global Equilibrium Asset Pricing Model with Home Preference. *Management Science*, 58(2), 273--292.

Tesar, L. L., and Werner, I. M. (1995). Home Bias and High Turnover. *Journal of International Money and Finance*, 14(4), 467--492.

Van Nieuwerburgh, S., and Veldkamp, L. (2009). Information Immobility and the Home Bias Puzzle. *Journal of Finance*, 64(3), 1187--1215.

Voronkova, S., and Bohl, M. T. (2005). Institutional Traders' Behavior in an Emerging Stock Market: Empirical Evidence on Polish Pension Fund Investors. *Journal of Business Finance and Accounting*, 32(7--8), 1537--1560.

### Companion Papers in This Series

Peters, B. (2026). Demographic Structure and International Capital Flows: A 140-Country Panel. Working Paper. [Paper 1]

Peters, B. (2026). Where Does Demographic Capital Go? A Bilateral Gravity Approach. Working Paper. [Paper 2]

Peters, B. (2026). Why FH Correlations Vary: Demographics and Savings Retention. Working Paper. [Paper 5]

Peters, B. (2026). Demographics and Asset Prices: The Murder-Suicide of the Rentier. Working Paper. [Paper 6]

Peters, B. (2026). Demographics and the Trilemma: How Population Aging Shapes Exchange Rate Regime Choice. Working Paper. [Paper 10]

Peters, B. (2026). Sample Composition Fragility: A Diagnostic Framework. Working Paper. [Paper 19]
